Click here to

Session: Deep Learning for Image-guided Therapy [Return to Session]

DECT-Based Material Mass Density Mapping for Proton Monte Carlo Dose Calculation Using Physics-Informed Deep Learning

C Chang1*, Y Gao2, T Wang3, Y Lei4, Q Wang5, J Bradley6, T Liu7, L Lin8, J Zhou9, X Yang10, (1) Emory University, Atlanta, GA, (2) Emory University, Atlanta, GA,(3) Emory University, Atlanta, GA, (4) Emory University, Atlanta, GA, (5) Emory University, Atlanta, GA, (6) Emory University School Of Medicine, Atlanta, GA,(7) Emory University, Atlanta, GA, (8) Emory Proton Therapy Center, Atlanta, GA, (9) Emory University, Atlanta, GA, (10) Emory University, Atlanta, GA


WE-C1030-IePD-F2-4 (Wednesday, 7/13/2022) 10:30 AM - 11:00 AM [Eastern Time (GMT-4)]

Exhibit Hall | Forum 2

Purpose: A framework of physics-informed deep learning (PIDL) based dual-energy CT (DECT) parametric mapping was developed to infer material characteristics for proton Monte Carlo dose calculation. This work investigates material conversion and dosimetric accuracy using PIDL models.

Methods: A PIDL framework was used to generate material mass density and relative stopping power (RSP) maps from DECT with artificial neural networks (ANN), physics-informed ANN (PANN), and physics-informed ResNet (PRN). The material characteristic maps acquired from empirical models (DECT) were compared to the PIDL method. Monte Carlo dose calculation (MCDC) was also performed for dosimetry comparisons using different parametric mapping methods. A CIRS adult anthropomorphic phantom (M701) was irradiated at an HN site using anterior multiple-single-energy proton beams ranging from 175 to 189 MeV with 1-MeV incremental energy. A MariXX PT was placed under a proton couch to measure the exit dose from the phantom. Gamma criteria of 3%/3mm were used for measurement/simulation comparisons.

Results: Mass density errors of bone by empirical model, ANN, PANN, and PRN are 3.0%, 3.8%, 2.5%, and 0.5%, and RSP errors are 3.2%, 3.5%, 2.5%, and 0.9%. The soft tissue mass density errors of soft tissues of the four models are 2.7%, 2.5%, 1.3%, and 1.1%, and RSP errors of lung are 7.5%, 11.0%%, 10.9%, and 0.7%, respectively. The mean Gamma passing rates are 94.8% and 98.5% for MCDC using mass density maps from the empirical model and PRN.

Conclusion: We demonstrated that the PIDL framework delivered accurate material mass density and RSP conversion using PRN and an anthropomorphic phantom. The dosimetry measurement verified that the PIDL framework improved the DECT parametric mapping and enhanced gamma agreement between MCDC and the experiment. The physics-informed training can reduce the uncertainty for mass density and RSP maps compared to conventional deep learning training (without constraints by a physics model).


Not Applicable / None Entered.


Not Applicable / None Entered.

Contact Email